01
Research paper
The Transformer paper replaces recurrence with attention
Google researchers introduce an attention-only architecture that trains efficiently in parallel and becomes the foundation of modern large language models.
CURATED TIMELINE · EDITORIAL EDITION
From Attention to Reasoning Models
Six releases map the dominant scaling story: the Transformer architecture, GPT-3’s in-context learning, ChatGPT’s mass-market interface, GPT-4’s multimodality, o1’s test-time compute, and DeepSeek-R1’s open-weight challenge.
Timeline overview
At a glance
Hover for the event briefScroll horizontally
6 nodes2017.06—2025.01 · Editorially curated · Updated 2026-07-15
Editorial thread
The timeline
Reviewed event briefs and original editorial context, ordered to show how the story changed over time.
- 02Research paper
GPT-3 demonstrates in-context learning at 175B parameters
OpenAI's 175B-parameter GPT-3 performs many tasks from instructions and a few examples in the prompt, without updating its weights.
- 03Product launch
OpenAI releases ChatGPT as a public research preview
ChatGPT turns instruction tuning and reinforcement learning from human feedback into an accessible conversational product for the general public.
- 04Model release
OpenAI launches the multimodal GPT-4
GPT-4 accepts image and text input, improves performance on complex instructions and professional exams, and becomes OpenAI's new product foundation.
- 05Model release
OpenAI previews o1, a model trained to spend more time reasoning
OpenAI releases o1-preview, shifting frontier-model competition toward reinforcement learning and additional computation at inference time.
- 06Open-source release
DeepSeek releases the open-weight reasoning model R1
DeepSeek publishes R1, R1-Zero, six distilled models, and its reinforcement-learning recipe, pushing open reasoning models into the frontier conversation.